Marta Pedretti, Cesar Alan Contreras, Rustam Stolkin, Alireza Rastegarpanah, Roberto Pagani
Intent-driven robotic disassembly requires translating operator language and noisy RGB-D perception into grounded, precedence-consistent removal plans. This work extends a perception-grounded large language model (LLM) planning framework for two-arm electric vehicle (EV) battery disassembly and evaluates the pipeline’s effectiveness after replacing the LLM backbone. The system converts live RGB-D detections into structured scene text, rewrites coordinates into a frame-consistent LUZD representation, generates an ordered removal sequence, validates it with an LLM verifier, and applies a deterministic anti-hallucination filter that removes labels or coordinates not present in the perceived scene. The replacement LLM is first evaluated within the prior ensemble-based pipeline. The upgraded system then combines a single-checkpoint generator-verifier workflow, revised generator and verifier prompts, and an interactive, inspectable human-in-the-loop planning interface that displays intermediate outputs and final filtered plans. On the same 200-scene, 600-intent benchmark used in earlier work, mean full-list correctness was 0.824 for the prior six-sample final pipeline, 0.896 when the replacement LLM was evaluated within that prior pipeline, 0.935 with the upgraded generator alone, and 0.959 with verification and filtering. The final upgraded configuration reduced the LLM-call count from seven to two. In a 12-participant within-subject pilot, mean time to accepted answer was lower on both tasks, the mean raw NASA-TLX score was lower, and all participants preferred the new interface. These results suggest that structured scene text, clear prompts, verifier-based correction, deterministic grounding, and inspectable interfaces enhance human-in-the-loop disassembly planning.
Branko Lukić, Miloš Petrović, Nikola Knežević, Zaviša Gordić, Arso M. Vukićević, Kosta Jovanović
Efficient packaging waste sorting is an important requirement for implementing circular economy strategies, yet many automated sorting solutions are designed for large-scale, high-throughput recycling facilities and are not well suited to small-batch industrial environments characterized by limited waste volumes, frequent material changes, and the need for rapid system reconfiguration. This paper presents the design, implementation, and laboratory validation of a modular collaborative robotic workstation for small-batch packaging waste sorting. The proposed system integrates a collaborative robot, RGB-D vision system, conveyor-based material flow, and a ROS2-based control architecture within a unified perception–decision–manipulation framework. The perception pipeline combines class-agnostic segmentation using the Segment Anything Model (SAM) with lightweight MobileNetV2-based classification for material recognition and is extended with optional modules for geometric reasoning and damage assessment. The proposed architecture reduces dependence on task-specific segmentation training and provides semantic information relevant to grasp planning, safety-aware task allocation, and human–robot collaboration. The workstation was validated under controlled laboratory conditions using three representative packaging waste categories: plastic bottles, aluminium cans, and Tetra Pak containers. The SAM + MobileNetV2 perception pipeline achieved an accuracy of 0.93, precision of 0.93, recall of 0.92, and F1-score of 0.92. System-level experiments demonstrated successful integration of RGB-D object localization, conveyor tracking, camera-to-robot coordinate transformation, and robotic pick-and-place execution. In contrast to conventional high-throughput sorting systems, the proposed workstation emphasizes modularity, flexibility, and collaborative operation for small-batch industrial packaging waste streams. In addition, workstation-level manipulation tests over 200 grasp attempts resulted in a 92.0% grasp success rate, an 89.0% end-to-end sorting success rate, and an estimated sorting rate of 5.73 picks/min under controlled conveyor-based conditions. The results demonstrate the feasibility of integrating foundation-model-based perception with collaborative robot in a modular workstation intended for small-batch industrial packaging waste sorting. The proposed approach emphasizes flexibility, modularity, and reduced reconfiguration effort for variable waste streams. The results provide a basis for further development toward adaptive and human–robot collaborative sorting, while broader validation under more diverse, cluttered, and industrially representative conditions remains necessary.
In human–human communication, encouragement has been closely associated with subjective well-being (SWB). However, the long-term effects of encouragement in human–robot interaction remain unclear, particularly in comparison with human-delivered encouragement and across different encouragement strategy types. SWB comprises both affective well-being (AWB) and cognitive well-being (CWB). This study investigated whether sustained verbal encouragement from a human partner and a conversational robot could enhance older adults’ SWB. A two-phase design was adopted. First, a preliminary online survey with 55 older Japanese adults (aged 65–74) was conducted to select encouragement strategies for the main experiment, identifying offering specific actions (high-efficacy) and distraction (low-efficacy) as two encouragement strategies that differed significantly in participants’ perceived effectiveness (p < 0.05). Based on this survey, 18 older Japanese adults (aged 62–90) were recruited for an exploratory 6-week longitudinal experiment. Participants interacted with either a human partner or a conversational robot once per week and were randomly assigned to one of four conditions: human partner or social robot (RoBoHoN) crossed with high- or low-perceived-efficacy encouragement strategy. SWB was assessed using questionnaires, while post-session open-ended responses, diary entries, and post-experiment interviews provided qualitative data on encouragement experience, perceived SWB changes, and participants’ perceived sources of SWB change. Results suggested that high-efficacy verbal encouragement may positively influence older adults’ SWB, possibly with a stronger effect than low-efficacy encouragement, with a crossover-like pattern emerging: for AWB, the advantage of high-efficacy over low-efficacy encouragement tended to be more pronounced in the robot condition (Δ gap = 2.8) than the human condition (Δ gap = 0.5), whereas for CWB, it appeared to be more pronounced in the human condition (Δ gap = 2.0) than the robot condition (Δ gap = 0.2); these crossover trends should be interpreted with caution given the small sample size. Qualitative findings further suggested that the conversation itself may have contributed to participants’ perceived emotional support in the Japanese cultural context. These preliminary longitudinal findings suggest that social robots and human partners may both contribute to supporting older adults’ SWB and may play complementary roles.
Human intelligence does not develop linearly: it follows a precise phase sequence. This article presents the hypothesis that this neurodevelopmental timeline corresponds to the dual-architecture artificial general intelligence (AGI) system previously proposed by the author, covering the 10 architectural components. We hypothesize that AGI development may be constrained by similar phase sequencing, address key objections (see Sections 3.1, 5.4, 5.5, and 5.6 for explicit objection-response pairs) through formal arguments, and provide indirect supporting evidence for six individual mechanisms on the MuJoCo platform: four-method anti-forgetting comparison: elastic weight consolidation (EWC), synaptic intelligence (SI), memory-aware synapses (MAS), and Naive. Curiosity-driven exploration, training stability analysis, multimodal cross-prediction, and warm-start routing analysis reveal negative transfer in genuinely novel scenarios, and a multi-novelty gradient experiment spanning five gravity levels (1.0×–2.0×) identifies a crossover threshold at approximately 1.25×–1.50× gravity on HalfCheetah-v5. An appendix provides a formal mathematical theory of token migration with two formal sketches and one implication.
Due to the complex structures, variable morphologies, and multiple feasible grasping strategies for entangled fabrics during stacked sorting, achieving stable robotic grasping in such scenarios remains highly challenging. This study formulates the stacked-fabric grasping task as a problem of structural perception and multi-attribute decision-making. A stacked-fabric grasping method is proposed. Specifically, a convex-structure perception method based on point cloud geometric features is developed to extract operable regions from point cloud data, and a set of candidate grasp poses is generated based on these structures. Furthermore, to address the difficulty of effectively evaluating candidate poses, a hierarchical fuzzy neural network is developed by simulating the human fuzzy decision-making process, enabling interpretable evaluation and selection of the optimal grasp pose. Comparative experiments are conducted on a constructed fabric grasping dataset. Compared with the best-performing baseline model, the proposed method improves the prediction accuracy of grasp feasibility, grasp separation effectiveness, and grasp reliability by 1.3%, 0.3%, and 0.7%, respectively. Robotic grasping experiments further show that the proposed method achieves a grasping success rate of 93.3% in randomly stacked scenarios. The results demonstrate that integrating convex structure perception with hierarchical fuzzy decision-making provides an effective and interpretable strategy for grasp pose localization and selection in stacked-fabric grasping.
